GloVe employs a squared loss function to fit precomputed global corpus statistics, such as word-word co-occurrence counts. This approach is chosen because cross-entropy loss is often suboptimal for measuring the difference between two probability distributions when dealing with an extremely large corpus.